lkml.org 
[lkml]   [2022]   [Feb]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH] KVM: x86/pmu: Distinguish EVENTSEL bitmasks for uniform event creation and filtering
From
On 10/2/2022 10:09 pm, Jim Mattson wrote:
> On Thu, Feb 10, 2022 at 2:26 AM Like Xu <like.xu.linux@gmail.com> wrote:
>>
>> From: Like Xu <likexu@tencent.com>
>>
>> The current usage of EVENTSEL_* macro is a mess in the KVM context. Partly
>> because we have a conceptual ambiguity when choosing to create a RAW or
>> HARDWARE event: when bits other than HARDWARE_EVENT_MASK are set,
>> the pmc_reprogram_counter() will use the RAW type.
>>
>> By introducing the new macro AMD64_EXTRA_EVENTSEL_EVENT to simplify,
>> the following three issues can be addressed in one go:
>>
>> - the 12 selection bits are used as comparison keys for allow or deny;
>> - NON_HARDWARE_EVENT_MASK is only used to determine if a HARDWARE
>> event is programmed or not, a 12-bit selected event will be a RAW event;
>> (jmattson helped report this issue)
>> - by reusing AMD64_RAW_EVENT_MASK, the extra 4 selection bits (if set) are
>> passed to the perf correctly and not filtered out by X86_RAW_EVENT_MASK;.
>>
>> Signed-off-by: Like Xu <likexu@tencent.com>
>> ---
>> arch/x86/include/asm/perf_event.h | 3 ++-
>> arch/x86/kvm/pmu.c | 11 ++++-------
>> arch/x86/kvm/pmu.h | 6 ++++++
>> 3 files changed, 12 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
>> index 8fc1b5003713..bd068fd19043 100644
>> --- a/arch/x86/include/asm/perf_event.h
>> +++ b/arch/x86/include/asm/perf_event.h
>> @@ -43,8 +43,9 @@
>> #define AMD64_EVENTSEL_INT_CORE_SEL_MASK \
>> (0xFULL << AMD64_EVENTSEL_INT_CORE_SEL_SHIFT)
>>
>> +#define AMD64_EXTRA_EVENTSEL_EVENT (0x0FULL << 32)
>> #define AMD64_EVENTSEL_EVENT \
>> - (ARCH_PERFMON_EVENTSEL_EVENT | (0x0FULL << 32))
>> + (ARCH_PERFMON_EVENTSEL_EVENT | AMD64_EXTRA_EVENTSEL_EVENT)
>> #define INTEL_ARCH_EVENT_MASK \
>> (ARCH_PERFMON_EVENTSEL_UMASK | ARCH_PERFMON_EVENTSEL_EVENT)
>>
>> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
>> index 2c98f3ee8df4..99426a8d7f18 100644
>> --- a/arch/x86/kvm/pmu.c
>> +++ b/arch/x86/kvm/pmu.c
>> @@ -198,7 +198,8 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
>>
>> filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
>> if (filter) {
>> - __u64 key = eventsel & AMD64_RAW_EVENT_MASK_NB;
>> + __u64 key = eventsel & (INTEL_ARCH_EVENT_MASK |
>> + AMD64_EXTRA_EVENTSEL_EVENT);
>>
>> if (bsearch(&key, filter->events, filter->nevents,
>> sizeof(__u64), cmp_u64))
>> @@ -209,18 +210,14 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
>> if (!allow_event)
>> return;
>>
>> - if (!(eventsel & (ARCH_PERFMON_EVENTSEL_EDGE |
>> - ARCH_PERFMON_EVENTSEL_INV |
>> - ARCH_PERFMON_EVENTSEL_CMASK |
>> - HSW_IN_TX |
>> - HSW_IN_TX_CHECKPOINTED))) {
>> + if (!(eventsel & NON_HARDWARE_EVENT_MASK)) {
>
> I still don't understand why we even bother doing this lookup in the
> first place. What's wrong with simply requesting PERF_TYPE_RAW every
> time?

Thanks for the constant chasing, I finally got a reply from Peterz:
"think so; the HARDWARE is just a convenience wrapper over RAW IIRC".

Let me take this step and clean it up a bit.

\
 
 \ /
  Last update: 2022-02-17 10:06    [W:0.048 / U:0.300 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site